Statistics Colloquium : Dr. Md. Siddiqur Rahman
Abstract: For building a linear prediction model, Robust version of Backward Elimination (RBE) is introduced in this study, which is computationally very suitable and scalable to large high-dimensional datasets. Since Backward Elimination (BE) can be expressed in terms of sample correlations, simple robustifications are obtained by replacing these correlations by their robust counterparts. In this article, the use of FastMCD (Fast Minimum Covariance Determinant) based correlation is proposed in BE to reduce the influence of outlying data points. We call this proposed method BEmcd. An extensive simulation study is conducted to compare the performance of BEmcd with RBE based on Winsorized correlation, Spearman rank correlation, Kendal’s Tau correlation and BE with Pearson correlation. Simulation studies and a real data application demonstrate the excellent performance of BEmcd.